Opportunities and Challenges for Scaling Up Robot Learning.
A long-standing vision of robotics has been the creation of autonomous robots that can learn from the world around them and assist humans in performing a variety of tasks including services in our homes, transportation and rescue in catastrophe-struck environments. However, most of the robots today are programmed to operate in carefully engineered factory settings or they are confined to only perform specific tasks in precisely modeled environments. This can be primarily attributed to the narrow scope of current methods that restrict their ability to quickly adapt to new tasks or previously unseen environments. In this talk, I will first present our state-of-the-art deep learning methods for several fundamental problems that are key enablers for robot autonomy including perception, localization, and prediction. I will then discuss recent advances that we are making in Freiburg for scaling up robot learning through self-supervision.